Face Image Matching Based on Improved Triplet Network
Xiaoshuo Jia, Zeng Shangyou, Chengxu Ma · 2019
On the basis of the Triplet Network, there are many limitations in using the Euclidean distance for similarity determination. This paper proposes an image preprocessing method NR (New Robinson), a lightweight My-net model and an NTn (New Triplet network) model of image relative dispersion DF (Dispersion Function) combination. In this paper, the characteristics of NR (New Robinson) are used to reduce the model size of CNN, and DF (Dispersion Function) is used instead of the traditional Euclidean distance algorithm to determine the similarity and reduce the generality of Euclidean distance. Experiments on two data sets VGG_FACE2 and CACD2000 show that the accuracy of NTn model is improved by 10.7% and 16.0% compared with traditional model; the size of model is reduced from 19.6 MB to 3.2 MB; DF decays faster than Euclidean distance on Loss trend map, and the oscillation degree is more stable. As can be seen from the results, NTn has more advantages in accuracy and model size than Triplet network model.